We broke the world's leading AI image detectors.
We benchmarked 7 commercial AI image detectors.
Basic edits like blur and noise dropped detection rates as low as 4%.
Edit this receipt image to exactly double the total amount and all intermediate monetary values, including line item prices, subtotal, tax amount, total, and any other amount necessary. Keep everything else in the photo exactly the same.
Instead of detecting what's fake, Prove What's Real.
AI image detectors are unreliable. Even the best commercial detectors can be defeated by simple post-processing, reducing detection rates from 90%+ to under 10%.
The answer isn’t to detect what’s fake. It’s to Prove What’s Real.
Proving what's real shifts the question from "is this AI?" to "can this be verified?"—a fundamentally stronger guarantee that doesn't degrade as generative models improve.
All 7 detectors failed across 15,630 images from 5 generative models
Product: We took real product photos and used AI to generate a new image with a realistic issue that a customer would complain about.
Show prompt
“Edit this photo so this product looks like it arrived from the seller with a realistic issue that a customer would complain about. The issue can be either minor shipping damage (scuff/tear/rip/dent, paint chip, or a small crack) OR a manufacturing defect (uneven paint finish, misaligned seam, or a surface blemish). It still has to look like a real photo.”
7 more ↓↑50%12%↓37
7 more ↓↑44%12%↓32
7 more ↓↑33%12%↓20
7 more ↓↑43%13%↓30
7 more ↓↑19%13%↓6
7 more ↓↑35%13%↓22
Showing 49–54 of 287 images